fix: gfx1151, wrong logits - #28574
Closed
Error00101 wants to merge 1 commit into
Closed
Conversation
Keep integrated GPU detection while staging gfx1151 buffers in device memory to avoid UMA prefill latency regressions. Co-authored-by: Cursor <cursoragent@cursor.com>
Author
|
Was fixed by d4389a4 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
ROCm Produces very high perplexity on RDNA3.5 with promts >n_ubatch (Reproduceable with llama-perplexity)
Disabling the host buffer foregoes the corruption.
Overview
From what i can read from the AMD Fork there is a MMQ tile-barrier race with the amd gfx115x series (host buffer to gpu).
AMD-Ecosystem#75
which then led to a bit of latency which is fixed by
AMD-Ecosystem@865374b
Now RDNA3.5 excludes direct host access while other architectures retain it.
Eliminates the whole race condition and subsequent corruption before it even happens.
Additional information
I did not create the fix. I just tested it localy and created this PR.
Fixes: Eval bug: HIP/ROCm on gfx1151, wrong logits (not a crash), triggered by prompts longer than n_ubatch.
#28211
Requirements
The fix was created by https://github.com/liangliangchang
I tested the fix: ROCm now produces the same perplexity as Vulkan